Deep Learning Model for Prediction of Mutational Profiles Associated with Clonal Hematopoiesis of Indeterminate Potential (CHIP) from Peripheral Blood Smears

Clonal Hematopoiesis of Indeterminate Potential (CHIP) is defined by the presence of clonal mutations in myeloid cells in the absence of cytopenias or hematologic malignancies. The prevalence of CHIP rises with age, reaching 10–20% by age 70, and CHIP is associated with a 0.5-1% risk of development of a hematological neoplasm, a tenfold increase compared to the general population. Multiple studies have reported higher incidences of CHIP in African American patients as well as a higher likelihood of developing CHIP-associated myeloid neoplasms.

CHIP has also been linked with a 40% increase in all-cause mortality, a twofold increase in the risk of coronary heart disease, an increased risk of myocardial infarction, aortic stenosis, stroke, worse outcomes in heart failure patients, higher risk of non-hematological cancers with poorer prognoses, as well as increased toxicities with oncologic therapies.

Detection of CHIP currently requires next-generation sequencing (NGS) of numerous target genes at an approximate cost ranging from $500-$2,000 per sample, rendering population-based screening impractical and contributing to disparities in the field.

Given the utility of deep-learning based characterization of cell images, we propose to develop an image analysis model trained using a multiple-instance learning framework capable of detecting CHIP using single-cell digitized images obtained from peripheral blood smears. If successful, this approach will provide a virtually cost-free screening method.

We are developing a deep-learning architecture to characterize single-cell peripheral blood smear images. We are currently constructing an internal database combining NGS results with single-cell images from patients with and without CHIP and hematologic malignancies. We will use this dataset to train a vision encoder-decoder architecture that can extract relevant features as embedding vectors profiling individual images/instances. Using the pretrained encoder network, we will develop and characterize a classifier to stratify CHIP/non-CHIP cases. The cell-level images from individual samples will be combined randomly to generate labeled 'bag' of individual cell images. Using this dataset, we will combine the image encoder with a MLP classifier and train using the multiple-instance learning framework. With the trained/established model, we will use saliency maps or integrated gradients to investigate morphological features relevant for classification at image/instance level.

In this preliminary phase, we have successfully built a model able to accurately identify different cell types based on images from peripheral blood smears. We are now training this model to differentiate between cells belonging to individuals with CHIP mutations and those that do not.

As most adults undergo routine basic blood testing and many clinical laboratories already employ software that digitized peripheral blood smears, this model will enable large-scale, population-based screening and facilitate the implementation of surveillance strategies and interventions aimed at reducing the downstream risks associated with CHIP.